This paper proposes a multilayer perceptron guided encryption system for secure data exchange in wireless communication, addressing key distribution challenges and man-in-the-middle attacks. The method ensures both sender and receiver generate identical session keys through synchronized weight vectors, enabling effective encryption and decryption of plaintext. The technique demonstrates comparable performance to existing classical methods under various test conditions.

1. International Journal of Mobile Network Communications & Telematics ( IJMNCT) Vol. 4, No.4,August 2014
SECURED TRANSMISSION THROUGH MULTI LAYER
PERCEPTRON IN WIRELESS COMMUNICATION
(STMLP)
Arindam Sarkar1 and J. K. Mandal2
1Department of Computer Science & Engineering, University of Kalyani, W.B, India
2Department of Computer Science & Engineering, University of Kalyani, W.B, India
ABSTRACT
In this paper, a multilayer perceptron guided encryption/decryption (STMLP) in wireless communication
has been proposed for exchange of data/information. Multilayer perceptron transmitting systems at both
ends generate an identical output bit and the network are trained based on the output which is used to
synchronize the network at both ends and thus forms a secret-key at end of synchronizations of the
networks. Weights or hidden units of the hidden layer help to form a secret session key. The plain text is
encrypted through chaining , cascaded xoring of multilayer perceptron generated session key. If size of the
final block of plain text is less than the size of the key then this block is kept unaltered. Receiver will use
identical multilayer perceptron generated session key for performing deciphering process for getting the
plain text. Parametric tests have been done and results are compared in terms of Chi-Square test, response
time in transmission with some existing classical techniques, which shows comparable results for the
proposed technique. Variation numbers of input vectors and hidden layers will increase the confusion
/diffusion of the schemeand hence increase the security. As a result variable energy based techniques may
be achieved which may be applicable devices/interface of the heterogeneous sizes of the network/device.
KEYWORDS
Multilayer Perceptron, Session Key, Wireless Communication.
1. INTRODUCTION
In recent times wide ranges of techniques are developed to protect data and information from
eavesdroppers [1, 2, 3, 4, 5, 6, 7, 8, 9]. Algorithms have their virtue and shortcomings. For
Example in DES, AES algorithms [1] the cipher block length is nonflexible. In NSKTE [4],
NWSKE [5], AGKNE [6], ANNRPMS [7] and ANNRBLC [8] technique uses two neural
network one for sender and another for receiver having one hidden layer for producing
synchronized weight vector for key generation. Attacker can get an idea about sender and
receiver’s neural machine as session architecture of neural machine is static. In NNSKECC
algorithm [9] any intermediate blocks throughout its cycle taken as the encrypted block and this
number of iterations acts as secret key. Here, if n number of iterations are needed for cycle
formation and if intermediate block is chosen as an encrypted block after n/2th iteration then
exactly same number of iterations i.e. n/2 are needed for decode the block which makes easier
the attackers life. This paper proposed a multilayer perceptron guided encryption technique in
wireless communication to overcome the problem.
The organization of this paper is as follows. Section 2 of the paper deals with the problem domain
and methodology. Proposed Multilayer Perceptron based key generation has been discussed in
DOI : 10.5121/ijmnct.2014.4401 1

2. International Journal of Mobile Network Communications & Telematics ( IJMNCT) Vol. 4, No.4,August 2014
section 3. Triangularization encryption technique is given in section 4. Triangularization
decryption has been presented in section 5. Section 6 presents energy computation technique.
Complexity analysis of the technique is given in section 7. Experimental results are described in
section 8. Analysis of the results presented in section 9. Analysis regarding various aspects of the
technique has been presented in section 10. Conclusions and future scope are drawn in section 11
and that of references at end.
2
2. PROBLEM DOMAIN AND METHODOLOGY
In security based communication the main problem is distribution of key between sender and
receiver. As, during exchange of key over public channel intruders can intercept the key as a
middleman. The problem has been addressed and a technique has been proposed addressing the
issue. These are presented in section 2.1 and 2.2 respectively.
2.1. Man-In-The-Middle Attack
Intruders intercepting in the middle between sender and receiver try to capture all the information
transmitting from both. Diffie-Hellman key exchange technique [1] suffers from this type of
problems. Intruders can act as sender/ receiver simultaneously and try to steal secret session key
at the time of exchanging key via public channel.
2.2. Methodology
Well known problem of man in the middle attack has been addressed in STMLP where secret
session key is not exchanged over public insecure channel. At end of synchronization both
parties’ generates identical weight vectors and activated hidden layer outputs for both the parties
become identical. This identical output of hidden layer for both parties are used as one time secret
session key for secured data exchange.
The basic idea here is to design such a program with effective GUI which helps people to
understand the underlying calculations. In this case this would be the Tree Parity Machine and the
various encryption and decryption techniques. First we need to figure out what are main functions
of our system. Since we are going to work on various Neural network structures we need a menu
to choose from. Again after that, two different Neural network need mutual synchronization and
associated statistical data like type of network and total time required to synchronize mutually.
Then at the end we need a menu to choose various encryption and decryption techniques and
statistical modules to compute probable power consumption by the network. So we need various
menus / forms to cater our need of various functions within their scope. So the schematic view
looks like the figure 1.

3. International Journal of Mobile Network Communications & Telematics ( IJMNCT) Vol. 4, No.4,August 2014
3
Figure 1.
Selection of Type
of Neural Network
Encryption
/Decryptio
Analysis
Module
K,N,L Tuning
Time
Data Base
Statistical Data
Figure 1. Schematic diagram
3.MULTILAYER PERCEPTRON BASED SESSION KEY
GENERATION
A multilayer perceptron synaptic simulated weight based undisclosed key generation is carried
out between recipient and sender. Figure 2 shows multilayer perceptron based synaptic simulation
system. Same single hidden layer among multiple hidden layers for a particular session. All other
hidden layers goes in deactivated mode with the incoming input. The key generation technique
with analysis using random number of nodes (neurons) along with the corresponding algorithm is
discussed in the subsections 3.1 to 3.5.
Figure 2. A multilayer perceptron with 3 hidden Layers
Multilayer perceptron in each session acts as a single layer network with dynamically chosen one
activated hidden layer and K no. of hidden neurons, N no. of input neurons having binary input
vector, Î{−1,+1} ij x , discrete weights, are generated from input to output, are lies between -L and
+L, w { L L L} ij Î − ,− +1,...,+ .Where i = 1,…,K denotes the ith hidden unit of the perceptron and j = 1,…,N
the elements of the vector and one output neuron. Output of the hidden units is calculated by the

4. International Journal of Mobile Network Communications & Telematics ( IJMNCT) Vol. 4, No.4,August 2014
weighted sum over the current input values . So, the state of the each hidden neurons is expressed
using (eq.1)
4
= = (1)
h w x
i i i i j,
i j
N
j
N
wx
N
1
,
1 1
=
Output of the ith hidden unit is defined as
sgn( ) i i s = h
(2)
In case of i h = 0, i s = -1 to produce a binary output. Hence, i s = +1, if the weighted sum over its
inputs is positive, or else it is inactive, i s = -1. The output of a perceptron is the product of the
hidden units expressed in (eq. 2).
t s (3)
Õ=
=
K
i
i
1
3.1 Simulation
Input:-Random weights, input vectors for both multilayer perceptrons.
Output:-Secret key through synchronization of input and output neurons as vectors. Method:-
Random vectors generated, fed into the networks. Vectors are updated only when output of
machines produce identical output . The process continue till both machines are fully
synchronized.
Step 1. Initialize of random weight values of synaptic links between input layer and
randomly selected activated hidden layer.
Where, { } L L L wi
j Î− ,− +1,...+, (4)
Step 2. Repeat step 3 to 6 until the full synchronization is achieved, using
Hebbian-learning rules.
( ( ) ( A B ))
i j i j i w g w x i j = + tQs t Qt t +
, , , (5)
Step 3. Generate random input vector X. Inputs are generated by a third party or one of the
communicating parties.
Step 4. Compute the values of the activated hidden neurons of activated hidden layer using
(eq. 6)
= = (6)
h w x
i i i i j,
i j
N
j
N
wx
N
1
,
1 1
=
Step 5. Compute the value of the output neuron using
t s (7)
Õ=
=
K
i
i
1
Compare the output values of both multilayer perceptron by exchanging the system
outputs.
if Output (A) Output (B), Go to step 3
else if Output (A) = Output (B) then one of the suitable learning rule is applied only
the hidden units are trained which have an output bit identical to the
common output.
Update the weights only if the final output values of the perceptron are equivalent. When
synchronization is finally achieved, the synaptic weights are identical for both the system.

5. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
5
i j i w g w f x i j , , , , , = + s t t +
Hebbian learning
anti-Hebbian learning
Random walk learning
3.2 Multilayer Perceptron Learning
At the beginning of the synchronization process multilayer perceptron of A and B start with
uncorrelated weight vectors A B
i w / . For each time step K, public input vectors are generated
randomly and the corresponding output bits t
A/Bare calculated. Afterwards A and B communicate
their output bits to each other. If they disagree, t
A
t
B, the weights are not changed. Otherwise
learning rules suitable for synchronization is applied. In the case of the Hebbian learning rule [10]
both neural networks learn from each other.
( ( ) ( )) A B
i j i j i wi j g w x = + tQst Qt t +
, , , (8)
The learning rules used for synchronizing multilayer perceptron share a common structure. That
is why they can be described by a single (eq. 4)
( ( A B
) ) i j
(9)
with a function ( A B )
i f s ,t ,t , which can take the values -1, 0, or +1. In the case of bidirectional
interaction it is given by
s
( ) ( ) ( )
, , =Q Q −
s
1
s t t st t t A B A A B
i f
(10)
The common part ( A ) ( A B ) Qst Qt t of ( A B )
i f s ,t ,t controls, when the weight vector of a hidden
unit is adjusted. Because it is responsible for the occurrence of attractive and repulsive steps [6].
3.3 Weight Distribution within Multilayer Perceptron
In case of the Hebbian rule (eq. 8), A's and B's multilayer perceptron learn their own output.
Therefore the direction in which the weight i j w , moves is determined by the product i i j x , s . As
the output i s is a function of all input values, i j x , and i s are correlated random variables. Thus
the probabilities to observe i i j x , s = +1 or i i j x , s = -1 are not equal, but depend on the value of the
corresponding weight i j w , [11, 13, 14, 15, 16].
1
w
( )
P s x erf (11)

8. i j
−
= = +
2
,
,
, 1
2
1
i i j
i i j
NQ w
According to this equation, sgn( ) i i, j i, j s x = w occurs more often than the opposite,
sgn( ) i i , j i , j s x = − w . Consequently, the Hebbian learning rule (eq. 8) pushes the weights
towards the boundaries at -L and +L. In order to quantify this effect the stationary probability

9. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
distribution of the weights for t ® ¥ is calculated for the transition probabilities. This leads to
[11].
6
( )
−
( )
Õ=

12. m
m
−
−

15. − −
+
= =
w
m
i
i
i j
NQ m
erf
NQ m
erf
P w w P
1
2
2
, 0
1
1
1
1
(12)
Here the normalization constant 0 r is given by
−
( )
1
1
2
2
0
1
1
1
1
−
L
= − =

24. −
−

27. m
− −
+
= Õ
w L
w
m
m
i
i
NQ m
erf
NQ m
erf
P
(13)
In the limit N ® ¥ the argument of the error functions vanishes, so that the weights stay
uniformly distributed. In this case the initial length of the weight vectors is not changed by the
process of synchronization.
( 1
)
3
( 0)
+
= =
L L
Q t i (14)
But if N is finite, the probability distribution itself depends on the order parameter i Q Therefore its
expectation value is given by the solution of the following equation:
L
( )
=−
= =
i i j Q w P w w ,
w L
2
(15)
3.4 Order Parameters
In order to describe the correlations between two multilayer perceptron caused by the
synchronization process, one can look at the probability distribution of the weight values in each
hidden unit. It is given by (2L + 1) variables.
P P(w a w b) B
i
= A
= Ù = a , b i , j
i , j
(16)
which are defined as the probability to find a weight with w a A
i j = , in A's multilayer perceptron and
w b B
i j = , in B's multilayer perceptron. In both cases, simulation and iterative calculation, the
standard order parameters, which are also used for the analysis of online learning, can be
calculated as functions of i
a b P ,
[12].
= − = −
1 2
A
= w w =
a P
i L
Q ,
a L
L
b L
i
a b
A
i
A
i
N
(17)
2 1
= − = −
B
= w w =
b P
i L
Q ,
a L
L
b L
i
a b
B
i
B
i
N
(18)
= − = −
1
AB
= w w =
abP
i L
R ,
a L
L
b L
i
a b
B
i
A
i
N
(19)

28. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
7
The level of synchronization is given by the normalized overlap between two corresponding
hidden units as
r = = (20)
B
i
AB
i
R
A
i
B
i
B
i
w w
A
i
A
i
B
i
A
AB i
i
Q Q
w w w w
3.5 Secret Session Key
At end of full weight synchronization process, weight vectors between input layer and activated
hidden layer of both multilayer perceptron systems become identical. Activated hidden layer’s
output of source multilayer perceptron is used to construct the secret session key. This session key
is not get transmitted over public channel because receiver multilayer perceptron has same
identical activated hidden layer’s output. Compute the values of the each hidden unit by

30. s sgn ( )
N
=
i ij ij w x
=
j
1
−
=
0
1
1
sgn x
if
if
if
0,
0,
0.
=
x
x
x
(21)
For example consider 8 hidden units of activated hidden layer having absolute value (1, 0, 0, 1, 0,
1, 0, 1) becomes an 8 bit block. This 10010101 become a secret session key for a particular
session and cascaded xored with recursive replacement encrypted text. Now final session key
based encrypted text is transmitted to the receiver end. Receiver has the identical session key i.e.
the output of the hidden units of activated hidden layer of receiver. This session key used to get
the recursive replacement encrypted text from the final cipher text. In the next session both the
machines started tuning again to produce another session key.
Identical weight vector derived from synaptic link between input and activated hidden layer of
both multilayer perceptron can also becomes secret session key for a particular session after full
weight synchronization is achieved.
4. ENCRYPTION
For encryption a triangular based technique has been described. During plain text encryption, in
the first phase consider a block S = s0
0 s0
1 s0
2 s0
3 s0
4 s0
5 … s0
n-2 s0
n-1 of size n bits, where s0
i = 0 or
1 for 0 = i = (n-1). Now, starting from MSB (s0
0) and the next-to-MSB (s0
1), bits are pair-wise
XORed, so that the 1st intermediate sub-stream S1 = s1
0 s1
1 s1
2 s1
3 s1
4 s1
5 … s1
n-2 is generated
consisting of (n-1) bits, where s1
j = s0
j Å s0
j+1 for 0 = j = n-2, Å stands for the exclusive OR
operation. This 1st intermediate sub-stream S1 is also then pair-wise XORed to generate S2 = s2
0
s2
1 s2
2 s2
3 s2
4 s2
5 … s2
n-3, which is the 2nd intermediate sub-stream of length (n-2). This process
continues (n-1) times to ultimately generate Sn-1 = sn-1
0, which is a single bit only. Thus the size of
the 1st intermediate sub-stream is one bit less than the source sub-stream; the size of each of the
intermediate sub-streams starting from the 2nd one is one bit less than that of the sub- stream
wherefrom it was generated; and finally the size of the final sub-stream in the process is one bit
less than the final intermediate sub-stream. Table 1 and figure 3 show the process.

31. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
8
Table 1
Options for choosing Target Block from Triangle
Option
Serial No.
Target Block Method of Formation
001
s0
0 s1
0 s2
0 s3
0 s4
0 s5
0 …
sn-2
0 sn-1
0
Taking all the MSBs starting
from the source block till the
last block generated
010
sn-1
0 sn-2
0 sn-3
0 sn-4
0 sn-5
0
… s1
0 s0
0
Taking all the MSBs starting
from the last block generated
till the source block
011
s0
n-1 s1
n-2 s2
n-3 s3
n-4 s4
n-5
… sn-2
1 sn-1
0
Taking all the LSBs starting
from the source block till the
last block generated
100
sn-1
0 sn-2
1 sn-3
2 sn-4
3 sn-5
4
… s1
n-2 s0
n-1
Taking all the LSBs starting
from the last block generated
till the source block
Option Serial No. 010 Option Serial No. 100
Option Serial No. 001 Option Serial No. 011
Figure 3. Options diagram for choosing Target Block from Triangle
Table 1describes different options for choosing target block from triangle. This option is
generated by modulo 4 division of the value of output neuron then adding 1. Then take the binary
version of the decimal no. Each block size is represented by 5 bits and 3 bits are used to denoting
the option no. for that block. So, total 8 bits are used to describe a single block length and option
chosen. For multiple blocks several 8bits are attached together preceded by first 8 bits (28 = 256
blocks can be formed in one session) to describe total no. of block to forms intermediate sub key.
Maximum length of this sub key will be (256 blocks X 8 bits per block) 2048 bits. This sub key is
padded in the front of the encrypted text. Then the multilayer perceptron generated synchronized
one time session key is repeatedly xored with the intermediate traingularized cipher text by
considering same key traingularized cipher text length. This mechanism is performed until all
the blocks get exhausted.
5. DECRYPTION
During decryption, the receiver’s multilayer perceptron generated synchronized one time session
key is xored with the cipher text. The technique of performing xoring is same that was in
encryption process. Finally from the outcomes intermediate encrypted block (E) and sub key
block is extracted and now key is use to decipher the E to get the source stream. To ease the
explanation of decryption technique, let us consider, e0
i-1 = si-1
n-i for 1 = i = n, so that the
encrypted block becomes E = e0
0 e0
1 e0
2 e0
3 e0
4 … e0
n-2 e0
n-1. After the formation of the triangle, for
the purpose of decryption, the block en-1
0 en-2
0 en-3
0 en-4
0 en-5
0 … e1
0 e0
0, i.e., the block constructed
by taking all the MSBs of the blocks starting from the finally generated single-bit block En-1 to E,
are to be taken together and it is to be considered as the decrypted block. Figure 4 show the

32. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
triangle generated and hence the decrypted block obtained. Here the intermediate blocks are
referred to as E1, E2, …, En-2 and the final block generated as En-1.
9
Option Serial No. 010 Option Serial No. 100
Option Serial No. 001 Option Serial No. 011
Figure 4. Generation of Source Block from Target
6. ENERGY VARIABILITY
The proposed schemes have a good potential of energy variability which can be incorporated and
may be adopted on the fly during transmission. We know that energy required for a blue tooth
communication is less than that of a WiFi communication. This variability of energy can be
incorporated into the encryption system through incorporating variable number of hidden layers
and input neurons. Table 2 shows the proposed network sizes for various types of wireless
networks.
Table 2. Network type vs size of neural network
Type of Network
Energy
Availability
Network Size and parameters
Wireless Personal Area
Networks (Bluetooth,
Infrared)
Very Low
No. of input layer neurons= 5 to 10
No. of Hidden layer neurons= 4 to 8
No. of Hidden layer = 1 to 3
Synaptic Depth (L)= +5 to -5
Wireless LAN
Low
No. of input layer neurons= 10 to 25
No. of Hidden layer neurons= 8 to 20
No. of Hidden layer = 3 to 4
Synaptic Depth (L)= +10 to -10
Wireless mesh network Medium
No. of input layer neurons= 25 to 40
No. of Hidden layer neurons= 20 to 35
No. of Hidden layer = 4 to 5
Synaptic Depth (L)= +15 to -15
Wireless MAN Moderate
No. of input layer neurons= 40 to 50
No. of Hidden layer neurons= 35 to 45
No. of Hidden layer = 5 to 6
Synaptic Depth (L)= +20 to -20
Wireless WAN Relatively High
No. of input layer neurons= 50 to 55
No. of Hidden layer neurons= 45 to 50
No. of Hidden layer = 5 to 6
Synaptic Depth (L)= +25 to -25
Cellular network High
No. of input layer neurons= 55 to 100
No. of Hidden layer neurons= 50 to 70
No. of Hidden layer = 5 to 6
Synaptic Depth (L)= +30 to -30

33. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
10
7. COMPLEXITY ANALYSIS
The complexity of the technique will be O(L), can be computed using following three steps.
Step 1. To generate a MLP guided key of length N needs O(N) Computational steps. The
average synchronization time is almost independent of the size N of the networks, at
least up to N=1000.Asymptotically one expects an increase like O (log N).
Step 2. Complexity of the encryption technique is O(L).
Step 2. 1. Triangularization encryption process takes O(L).
Step 2. 2. MLP based encryption technique takes O(L) amount of time.
Step 3. Complexity of the decryption technique is O(L).
Step 3. 1. In MLP based decryption technique, complexity to convert final cipher
text into Tra cipher text T takes O(L).
Step 3. 2. Transformation of cipher text T into the corresponding stream of bits S
= s0 s1 s2 s3 s4…sL-1, which is the source block takes O(L) as this step also takes
constant amount of time for merging s0 s1 s2 s3 s4…sL-1.
So, overall time complexity of the entire technique is O(L).
8. RESULTS
In this section the results of implementation of the proposed STMLP technique has been
presented in terms of encryption decryption time, Chi-Square test, source file size vs. encryption
time along with source file size vs. encrypted file size. The results are also compared with
existing RSA [1] technique, existing ANNRBLC[8] and NNSKECC[9].
Table 3. Encryption / decryption time vs. File size
Encryption Time Decryption Time
Source
Size(bytes)
STMLP NNSKECC[9]
Encrypted
Size(bytes)
STMLP NNSKECC[9]
18432 6. 42 7.85 18432 6.99 7.81
23044 9. 23 10.32 23040 9.27 9.92
35425 14. 62 15.21 35425 14. 47 14.93
36242 14. 72 15.34 36242 15. 19 15.24
59398 25. 11 25.49 59398 24. 34 24.95
Table 3 shows encryption and decryption time with respect to the source and encrypted size
respectively. It is also observed the alternation of the size on encryption.
In figure 5 stream size is represented along X axis and encryption/decryption time is represented
along Y-axis. This graph is not linear, because of different time requirement for finding
appropriate MLP key. It is observed that the decryption time is almost linear, because there is no
MLP key generation process during decryption.

34. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
Chi-Square value
11
Encryption decryption time
9.23
Encryption Decryption
18432 23044 35425 36242 59398
Source size
14.62 14.72
9.27
14.47 15.19
Figure 5 Encryption decryption time against stream size
6.42
25.11
6.99
24.34
30
25
20
15
10
5
0
Table 4 shows Chi-Square value for different source stream size after applying different
encryption algorithms. It is seen that the Chi-Square value of STMLP is better compared to the
algorithm ANNRBLC [8] and comparable to the Chi-Square value of the RSA algorithm.
Table 4. Source size vs. Chi-Square value
Stream Size
(bytes)
Chi-Square value
(TDES) [1]
Chi-Square value in
(STMLP)
1500 1228.5803 2856.2673 2471.0724 5623.14
2500 2948.2285 6582.7259 5645.3462 22638.99
3000 3679.0432 7125.2364 6757.8211 12800.355
3250 4228.2119 7091.1931 6994.6198 15097.77
3500 4242.9165 12731.7231 10572.4673 15284.728
Figure 6 shows graphical representation of table 4.
Chi-Square value
(ANNRBLC) [8]
Figure 6. Chi-Square value against stream size
(RSA) [1]
Table 6 shows total number of iteration needed and number of data being transferred for MLP
key generation process with different numbers of input(N) and activated hidden(H) neurons and
varying synaptic depth(L).

35. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
Table 5. Data Exchanged
. and No. of Iterations For Different Parameters Value
Synaptic
Weight (L)
Total No. of
Iterations
Following figure 7. Shows the snapshot of MLP key simulation process. This snapshot shows the
tunning process of two multilayer perceptron with 4 hidden neurons, 4 input neurons and 6 as a
weight value with hebbian learning rule.
Figure 7.
Figure 8 shows the encryption and decryption time of a .txt file. File size taking as a bytes and
encryption/ decryption time represnts as a nanosecond.
Figure 8.
No. of
Input
Neurons(N)
No. of Activated
Hidden Neurons(K)
5 15
30 4
25 5
20 10
8 15
MLP Key Simulation Snapshot
Snapshot of Encryption and decryption time
Data
Exchanged
(Kb)
3 624 48
4 848 102
3 241 30
3 1390 276
4 2390 289
12

36. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
Figure 9 shows the memory heap representation of the key generation technique. The violet color
area represents the memory that has alreay been allocated. Another color represents the total
available memory.
13
Figure 9. Memory Map for whole application
Figure 10 shows the memory allocation gantt chart during key generation process.
Figure 10. Memory Gantt Chart
9. ANALYSIS
From results obtained it is clear that the technique will achieve optimal performances. Encryption
time and decryption time varies almost linearly with respect to the block size. For the algorithm
presented, Chi-Square value is very high compared to some existing algorithms. A user input key
has to transmit over the public channel all the way to the receiver for performing the decryption
procedure. So there is a likelihood of attack at the time of key exchange. To defeat this insecure
secret key generation technique a neural network based secret key generation technique has been
devised. The security issue of existing algorithm can be improved by using MLP secret session
key generation technique. In this case, the two partners A and B do not have to share a common
secret but use their indistinguishable weights or output of activated hidden layer as a secret key
needed for encryption. The fundamental conception of MLP based key exchange protocol focuses
mostly on two key attributes of MLP. Firstly, two nodes coupled over a public channel will
synchronize even though each individual network exhibits disorganized behaviour. Secondly, an
outside network, even if identical to the two communicating networks, will find it exceptionally
difficult to synchronize with those parties, those parties are communicating over a public
network. An attacker E who knows all the particulars of the algorithm and records through this
channel finds it thorny to synchronize with the parties, and hence to calculate the common secret
key. Synchronization by mutual learning (A and B) is much quicker than learning by listening (E)
[10]. For usual cryptographic systems, we can improve the safety of the protocol by increasing of

37. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
the key length. In the case of MLP, we improved it by increasing the synaptic depth L of the
neural networks. For a brute force attack using K hidden neurons, K*N input neurons and
boundary of weights L, gives (2L+1)KN possibilities. For example, the configuration K = 3, L =
3 and N = 100 gives us 3*10253 key possibilities, making the attack unfeasible with today’s
computer power. E could start from all of the (2L+1)3N initial weight vectors and calculate the
ones which are consistent with the input/output sequence. It has been shown, that all of these
initial states move towards the same final weight vector, the key is unique. This is not true for
simple perceptron the most unbeaten cryptanalysis has two supplementary ingredients first; a
group of attacker is used. Second, E makes extra training steps when A and B are quiet [10]-[12].
So increasing synaptic depth L of the MLP we can make our MLP safe.
14
10. SECURITY ISSUE
The main difference between the partners and the attacker in MLP is that A and B are able to
influence each other by communicating their output bits A t B t while E can only listen to these
messages. Of course, A and B use their advantage to select suitable input vectors for adjusting the
weights which finally leads to different synchronization times for partners and attackers.
However, there are more effects, which show that the two-way communication between A and B
makes attacking the MLP protocol more difficult than simple learning of examples. These
confirm that the security of MLP key generation is based on the bidirectional interaction of the
partners. Each partener uses a seperate, but identical pseudo random number generator. As these
devices are initialized with a secret seed state shared by A and B. They produce exactly the same
sequence of input bits. Whereas attacker does not know this secret seed state. By increasing
synaptic depth average synchronize time will be increased by polynomial time. But success
probability of attacker will be drop exponentially Synchonization by mutual learning is much
faster than learning by adopting to example generated by other network. Unidirectional learning
and bidirectional synchronization. As E can’t influence A and B at the time they stop transmit due
to synchrnization. Only one weight get changed where, = T. So, difficult to find weight for
attacker to know the actual weight without knowing internal representation it has to guess.
11. FUTURE SCOPE CONCLUSION
i s
i s
This paper presented a novel approach for generation of secret key proposed algorithm using
MLP simulation. This technique enhances the security features of the key exchange algorithm by
increasing of the synaptic depth L of the MLP. Here two partners A and B do not have to
exchange a common secret key over a public channel but use their indistinguishable weights or
outputs of the activated hidden layer as a secret key needed for encryption or decryption. So
likelihood of attack proposed technique is much lesser than the simple key exchange algorithm.
Future scope of this technique is that this MLP model can be used in wireless communication.
Some evolutionary algorithm can be incorporated with this MLP model to get well distributed
weight vector.
ACKNOWLEDGEMENT
The author expresses deep sense of gratitude to the DST, Govt. of India, for financial assistance
through INSPIRE Fellowship leading for a PhD work under which this work has been carried out.

38. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
15
REFERENCES
[1] Atul Kahate, Cryptography and Network Security, 2003, Tata McGraw-Hill publishing Company
Limited, Eighth reprint 2006.
[2] Sarkar Arindam, Mandal J. K, “Artificial Neural Network Guided Secured Communication
Techniques: A Practical Approach” LAP Lambert Academic Publishing ( 2012-06-04 ), ISBN:
978-3-659-11991-0, 2012
[3] Sarkar Arindam, Karforma S, Mandal J. K, “Object Oriented Modeling of IDEA using GA based
Efficient Key Generation for E-Governance Security (OOMIG) ”, International Journal of
Distributed and Parallel Systems (IJDPS) Vol.3, No.2, March 2012, DOI :
10.5121/ijdps.2012.3215, ISSN : 0976 - 9757 [Online] ; 2229 - 3957 [Print]. Indexed by: EBSCO,
DOAJ, NASA, Google Scholar, INSPEC and WorldCat, 2011.
[4] Mandal J. K., Sarkar Arindam, “Neural Session Key based Traingularized Encryption for Online
Wireless Communication (NSKTE)”, 2nd National Conference on Computing and Systems,
(NaCCS 2012), March 15-16, 2012, Department of Computer Science, The University of
Burdwan, Golapbag North, Burdwan –713104, West Bengal, India. ISBN 978- 93-808131-8-9,
2012.
[5] Mandal J. K., Sarkar Arindam, “Neural Weight Session Key based Encryption for Online
Wireless Communication (NWSKE)”, Research and Higher Education in Computer Science and
Information Technology, (RHECSIT- 2012) ,February 21-22, 2012, Department of Computer
Science, Sammilani Mahavidyalaya, Kolkata , West Bengal, India. ISBN 978-81- 923820-0-
5,2012
[6] Mandal J. K., Sarkar Arindam, “An Adaptive Genetic Key Based Neural Encryption For Online
Wireless Communication (AGKNE)”, International Conference on Recent Trends In Information
Systems (RETIS 2011) BY IEEE, 21-23 December 2011, Jadavpur University, Kolkata, India.
ISBN 978-1-4577-0791-9, 2011
[7] Mandal J. K., Sarkar Arindam, “An Adaptive Neural Network Guided Secret Key Based
Encryption Through Recursive Positional Modulo-2 Substitution For Online Wireless
Communication (ANNRPMS)”, International Conference on Recent Trends In Information
Technology (ICRTIT 2011) BY IEEE, 3-5 June 2011, Madras Institute of Technology, Anna
University, Chennai, Tamil Nadu, India. 978-1-4577-0590-8/11, 2011
[8] Mandal J. K., Sarkar Arindam, “An Adaptive Neural Network Guided Random Block Length
Based Cryptosystem (ANNRBLC)”, 2nd International Conference on Wireless Communications,
Vehicular Technology, Information Theory And Aerospace Electronic System Technology”
(Wireless Vitae 2011) By IEEE Societies, February 28- March 03, 2011,Chennai, Tamil Nadu,
India. ISBN 978-87-92329-61-5, 2011
[9] Mandal J. K., Sarkar Arindam, “Neural Network Guided Secret Key based Encryption through
Cascading Chaining of Recursive Positional Substitution of Prime Non-Prime (NNSKECC)”,
International Confference on Computing and Systems, ICCS – 2010, 19–20 November,
2010,Department of Computer Science, The University of Burdwan, Golapbag North, Burdwan –
713104, West Bengal, India.ISBN 93-80813-01-5, 2010
[10] R. Mislovaty, Y. Perchenok, I. Kanter, and W. Kinzel. Secure key-exchange protocol with an
absence of injective functions. Phys. Rev. E, 66:066102,2002.
[11] A. Ruttor, W. Kinzel, R. Naeh, and I. Kanter. Genetic attack on neural cryptography. Phys. Rev. E,
73(3):036121, 2006.
[12] A. Engel and C. Van den Broeck. Statistical Mechanics of Learning. Cambridge University Press,
Cambridge, 2001.
[13] T. Godhavari, N. R. Alainelu and R. Soundararajan “Cryptography Using Neural Network ” IEEE
Indicon 2005 Conference, Chennai, India, 11-13 Dec. 2005.gg
[14] Wolfgang Kinzel and ldo Kanter, Interacting neural networks and cryptography, Advances in
Solid State Physics, Ed. by B. Kramer (Springer, Berlin. 2002), Vol. 42, p. 383 arXiv- cond-mat/
0203011, 2002
[15] Wolfgang Kinzel and ldo Kanter, Neural cryptography proceedings of the 9th international
conference on Neural Information processing(ICONIP 02).h
[16] Dong Hu A new service based computing security model with neural
cryptographyIEEE07/2009.J

39. International Journal of Mobile Network Communications Telematics ( IJMNCT) Vol. 4, No.4,August 2014
16
Arindam Sarkar
INSPIRE FELLOW (DST, Govt. of India), MCA (VISVA BHARATI, Santiniketan,
University First Class First Rank Holder), M.Tech (CSE, K.U, University First Class
First Rank Holder). Total number of publications 25.
Jyotsna Kumar Mandal
M. Tech.(Computer Science, University of Calcutta), Ph.D.(Engg., Jadavpur
University) in the field of Data Compression and Error Correction Techniques,
Professor in Computer Science and Engineering, University of Kalyani, India. Life
Member of Computer Society of India since 1992 and life member of cryptology
Research Society of India. Dean Faculty of Engineering, Technology Management, working in the
field of Network Security, Steganography, Remote Sensing GIS Application, Image Processing. 25
years of teaching and research experiences. Eight Scholars awarded Ph.D. and 8 are pursuing. Total
number of publications 267.